Algorithmic Guardrails for Echocardiography AI

BACKGROUND: Automated echocardiography analysis using artificial intelligence (AI) currently lacks effective guardrails or oversight mechanisms to ensure its safe and reliable deployment. OBJECTIVES: This study investigated whether deep feature representations can identify echocardiograms at risk of misinterpretation by echocardiography AI-based left ventricular ejection fraction (LVEF) estimation models. METHODS: We retrospectively validated 2 state-of-the-art models for LVEF estimation, EchoNet (single-task model) and PanEcho (multitask foundation model), on 4,010 apical-4-chamber videos from 1,353 studies (n = 1,208) with biplane LVEF measurements in the MIMIC-IV-ECHO database (2017-2019). We extracted 4 deep feature representations from a view classifier, a left ventricle segmentation model, and the 2 validated LVEF estimation models. These features were evaluated for their ability to identify systematic misinterpretations (error slicing) and automatically exclude AI estimates for high-risk inputs (selective estimation). RESULTS: : -0.07 [-0.11, -0.03]), both models systematically overestimated LVEF at lower values (proportional bias slope, EchoNet: -0.33 [-0.39, -0.28], P < 0.001, PanEcho: -0.31 [-0.36, -0.26], P < 0.001). All investigated deep feature representations identified error slices for both models. Shared error slices between the models revealed 2 common failure modes: contrast-enhanced echocardiograms and poor-quality apical-4-chamber views. Selective estimation using intermediate features from EchoNet achieved the best performance across both estimation models (area under risk-coverage curve, EchoNet: 5.04 [4.45, 5.66]; PanEcho: 5.21 [4.54, 5.92]). CONCLUSIONS: Deep feature representations can be leveraged to identify and automatically flag misinterpretation-prone echocardiograms, enabling promising guardrail mechanisms for the reliable deployment of echocardiography AI.